The purpose of this study was that to investigate the consequence of missing data in nonuniform differential item functioning (DIF) detection. The independent variables were sample size (focal/reference: 250/250, 500/500, and 1000/1000), impact (mean of latent trait for reference vs. focal groups: 0 vs. 0, -0.5 vs. 0, and 0.5 vs. 0), percentage of missing data (0%, 10%, 20%, and 30%), magnitude of DIF (0, 0.4, 0.8, and 1), and type of missing data (missing completely at random, two types of missing at random and missing not at random). Five methods were used for dealing with missing data: listwise deletion, zero imputation, multiple imputation, stochastic regression imputation and complete (base line). Three nonuniform DIF detection methods were used: logistic regression, crossing SIBTEST, and IRTLR. The dependent variables were type I error rates, power and effect size.
The results shown that, first, the zero imputation method was the worst method for nonuniform DIF detection. Second, the listwise deletion method produced the similar results to the complete method. Third, the multiple imputation method appeared to preferable to stochastic regression imputation.
Comments & Questions:
1. Whether the results from the listwise deletion are still similar to those from the complete method when the percentage of missing data is manipulated as a higher percentage, such as 40% or 60%? That is because the author mentioned that the listwise deletion will reduce the effect size when large amount of missing data is exist, and in turn leads to a reduction in statistical power for hypothesis testing.